Algorithmic trading and AI

In the following section our attention turns to the literature analysis of research papers published on both algorithmic trading and the usage of AI for trading purposes. Starting with the research paper by Paiva et al. (2019) we are able to observe some findings re-lating to the forecasting ability of AI. The authors detail the inherent complexity of the process of price and return forecasting for the stock market which is caused by the na-ture of the market itself. It is especially the dynamics of market prices and the large amount of the so-called noise within those prices that makes it difficult to detect what factors are truly meaningful for the process of forecasting. Additionally, the market is impacted by various external factors on a continuous basis, making the stock market an incredibly complex playground for different types of models.

The authors detail this further by breaking down the process of forecasting as one in-volving only the linkage between the past and the present. Especially interesting is their discussion of the two main methods used both in the literature and on the field for this process. The first one being different econometric models based on statistics and imple-mented using trading algorithms and the second one being the usage of advanced ML models which are then implemented by the means of AI.

The trading algorithms are defined as ones using tools such as linear regression and GARCH-modelling. Whereas AI algorithms are noted as using artificial neural networks, random forests, support vector machines and other similar frameworks for their process of generating forecasts. Also, the level of flexibility of AI models is detailed as the authors describe the ability of these models to utilize both quantitative and qualitative sources of data. While trading algorithms need to rely mostly on financial time series data, AI models are able to function with a much more flexible and diverse dataset and work with data that is imperfect.

In their research paper, AI models are based on technical analysis, meaning that they are functioning based on return data for individual securities. The authors note that when working on the same data, AI models as opposed to trading algorithms are able to find complex patterns and so-called hidden meaning behind the data, which refers to com-plex relationships and causalities that would otherwise be impossible to detect. Also, the main reasoning behind using technical analysis is detailed as they note that this revolves around the belief that past patterns in prices repeat themselves and hence, that prices do not follow a random walk process.

In their final findings, the authors are able to show that their AI model is able to generate meaningful and significant returns, but the authors also note the great impact that trans-action costs have on the profitability that their model is able to obtain. When more re-alistic assumptions are taken into play and transaction costs are accounted for, the model struggles to make a profit.

Dash and Dash (2016) on the other hand detail a lot of background information regarding the usage of algorithmic trading and AI. They especially note the increasing relevance of the topic as data is currently more available than ever before and this makes it possible to develop highly advanced models. Similarly, to the paper by Paiva et al. (2019), the authors describe the forecasting based on financial time series data as a very difficult process, noting the different trends, variations and irregularities within the data.

While the difficulty of understanding this data is being understood, they also deem AI models as best suited for this purpose due to their high level of automation, speed and flexibility for going through these very large datasets and finding hidden meaning. The process of data mining is therefore also mentioned, and this is described as simply in-volving the extraction of meaningful statistics from big data as detailed by Witten et al.

(2011, p. 191-202).

AI models are seen as both tools for automating and as advantages for decision making.

As more information is available to investors using AI, Dash and Dash (2016) note that this will likely also enable a reduction in the level of risk that the investor needs to take in order to obtain a profit. When it comes to literature in the field, the authors note the common trend of using technical analysis to create different types of indicators, which are then used to develop trading signals and strategies for or by the models to generate returns.

The authors also detail the use of supervised AI models, which entail the training of these models using a set of inputs along with a set of desired outputs. The process of trading is trained as a simple classification task, where the buy, sell and hold decisions simply represent a set of outputs based on some set of inputs. In their final conclusions, the authors do not deem the sole use of technical analysis as being sufficient as they note the need for the usage of different types of big data analysis to further the probability of their AI forecasts.

White (2000) discusses some pitfalls of the datamining approach as he notes that some perceived results are only caused by luck instead of real forecasting ability. Gerlein et al.

(2016) on the other hand note the unbiasedness of AI model creation guaranteed by the splitting of the data into so called training sets and testing sets. With this approach the AI model is first trained on the training set and later the actual forecasting ability of the model is validated by applying it on the testing set that it has not been exposed to before.

Continuing on the research paper by Gerlein et al. (2016) focused on the creation of profitable ML algorithms, the authors similarly note that AI is well suited for the process of finding hidden relationships within data and consequently having a strong capability towards forecasting. It is also noted that most research papers and experts in the field train their AI models based on different types of variables, attributes and indicators that have been processed from the financial time series data, instead of using this raw data on its own.

The authors note that while the usage of AI models seem to imply better forecasts, this does not always translate to higher profits. In this regard, they note that different models must always be evaluated based on their actual performance and not solely based on the accuracy of their predictions, as this does not always reflect well when being applied to actual live markets. Especially higher volatility situations are seen as troublesome for AI models due to the fact that this renders the generalization of forecasts and finding causalities increasingly difficult. It is noted that traders should incorporate the results of multiple AI models as a weighted average to obtain meaningful results.

Matias and Reboredo (2012) further this discussion of AI models by noting their ad-vantages in solving different types of problems by using nonlinear data. Ballings et al.

(2015) additionally detail that various factors influence the stock market, and this results in highly nonlinear price data for the market as a whole. Hence, it can be seen that this nonlinearity an important aspect to consider as far as predictions are concerned and again the suitability of AI models for the purpose of price predictions is displayed.

Mullainathan and Spiess (2017) uncover more how AI models function and also discuss what types of developments have led to their creation. Firstly, they discuss both the con-tribution created by advances in computing and the findings made in the field of statis-tics. The way these models function is simply explained by means of comparison against standard algorithms which need distinct rulesets to carry out their tasks. AI models on the other hand are given an input and an output and the models are tasked with finding an underlying function that best predicts an output based on a set of inputs.

As such, AI and ML have a lot more freedom in finding different solutions and therefore also the results might bring additional findings that were not originally considered. The authors especially note that AI is not to be used only to solve old problems using new ways, but to solve completely new problems, before thought too difficult to tackle. Due to the noted ability of AI in finding hidden patterns, one can also hypothesize that it would be very well suited in the field of technical analysis where the uncovering of dif-ferent patterns within the return data are used as trading signals.

The authors note this ability of AI as these models are able to discover patterns without the need for specifying them in advance. In regard to typical regressions, AI models are noted as finding optimal models of best fit especially in nonlinear datasets. Also, the before mentioned flexibility of AI is noted in the research paper and these models are explained simply as tools to extract substance from big data.

Finally, they note that AI models are allowed to choose the models and rules that work best for the data and no such rules are specifically programmed. Therefore, these models find meaning based on the data itself and not on the presumptions of the programmer, making them likely less to be biased and better suited for the task at hand.

Antweiler and Frank (2004) for one create a ML algorithm to go through online posts as a way to forecast and explain stock market volatility and they are able to obtain a statis-tically significant small positive performance. Hendry and Clements (2004) additionally

make interesting findings noting that combining multiple forecasts from multiple differ-ent models creates more accurate results than simply relying on one model, a view which is shared by Bates and Granger (1969).

Wolff and Neugebauer (2019) set out to study how well different types of ML models are suited for stock return predictions, noting the wide use and acceptance of these models in other fields such as facial recognition. They also define AI models by their ability to learn as opposed to static rule-based algorithms and similarly to past studies they also note how well the models are suited for nonlinear datasets. Interestingly, the authors are still unable to find significant outperformance by these models as opposed to more advanced types of linear regressions. When it comes to the process of stock return predicting clear outperformance is still noted against a buy-and-hold strategy.

In the research paper the training of the model is seen as of particular importance along with the usage of new data to test the model on to obtain unbiased results. It is also seen that the models need a large amount of data during the training phase to obtain decent forecasts in live environments. One key observation is the ability of the program-mer to control certain tuning parameters for an AI model which in term determine the degree of fit that the model will aim for. While a model can be almost perfectly fit to the training data, this in term results in poor performance when the same model is exposed to out-of-sample data due to overfitting. This is also noted to be of concern when dealing with stock market predictions, as there is such a large amount of noise within the data.

Also, different complexities of AI models are examined along with their pros and cons, and the problem noted with very complex models is the large amount of training data that they need. Conversely, these complex models are also described to be specially well suited to model complex relationship as they have inherent flexibility. The authors detail that this need for data becomes a problem when using solely financial time series data due to the noise and changing factors that drive returns over time. Therefore, older re-turn data is significantly less relevant.

Lastly, it is noted that while ML and AI technologies are beginning to be more widely used within the asset management industry, the low signal-to-noise ratio of stock return data makes advanced linear models the preferred option. Still they note that AI models are better when the number of potential predictors for forecasting is very large within the dataset in question.

Treleaven et al. (2013) examine a lot of descriptive information on the usage of algorith-mic trading and similarly to the hedge fund industry, the secretive nature of the field is uncovered. In their study algorithmic trading simply refers to the usage of algorithms to automate either any part or the entirety of the trading process. In terms of hedge funds using algorithmic trading the real-life implementation process of this trading style is also detailed, with pre-trade analysis, signal generation, trade execution, post-trade analysis, risk management and asset allocation noted as key steps by the authors.

For the process of both creating and improving models, especially backtesting and dif-ferent simulations based on historical data are seen as relevant. Additionally, the risk of employing these systems is noted, with possible programming errors resulting in unex-pected behaviors and great potential losses. Some of the main challenges for both the implementation and literature within the field are noted as being the lack of understand-ing of the interactions that these algorithms have amongst each other and the widely varying behaviors that these systems exhibit if certain variables are changed.

Khandani and Lo (2011) analyze situations where different trading algorithms used by systematic traders are seen as exhibiting a high degree of correlation amongst each other.

The authors note that similar factors are used by fund managers as they try to take ad-vantage of identical anomalies. This in term can lead to the unwind hypothesis during market downturns, where the commonality of these traders creates a race to the bottom where losses are seen across the trading style.

Dawes (1979) finds that algorithms are better forecasters than humans and Grove et al.

(2000) mirror this view noting how algorithms are able to show outperformance against their human counterparts when it comes to forecasting. Arkes et al. (1986) interestingly note that while algorithms are better, as people obtain more experience within a field their usage of these systems is reduced, leading to worse performance.

Shaffer et al. (2013) also note that the reliance towards algorithms is seen as a negative in some fields and Promberger and Baron (2006), Önkal et al. (2009), Diab et al. (2011) and Eastwood et al. (2012) all show that people have an inherent preference towards predictions made by humans as opposed to those created by algorithms.

Dietvorst et al. (2015) continue on their findings, noting that some people showcase genuine algorithm aversion, which they deem as a mistrust of systems. While algorithms are noted as performing notably better than human forecasters, the forecasting output by humans is preferred. Interestingly, they are also able to show that an algorithm is judged more harshly if it makes an error as opposed to a human, even if the resulting financial consequences of the mistakes made by algorithms are notably smaller.

The authors detail that while there is a general consensus that algorithms are able to avoid small mistakes such as typos due to their automated processes, humans often deem algorithms and AI as unable to learn from experience and mistakes, hence ending up preferring human forecasters. Therefore, when making the same mistake, algorithms and humans are not evaluated equally. This is something that is noted to be counterpro-ductive as algorithms are superior and can produce significant additional value.

Box et al. (2015, p. 2-16) and Engle (1982) on the other hand show that algorithmic trad-ing has been carried out through the usage of statistical models in the past and Chen et al. (2006), Li and Kuo (2008) and Tenti (1996) show that while AI and ML models are superior in other fields, their performance in trading is disputed. Lastly, Kim (2003) shows that different types of models are required for different types of market

conditions and more importantly that frequent retraining of these AI models is needed in order to maintain their forecasting accuracy. This in term is due to the dynamic and continuously changing nature of the stock market, where the factors driving returns evolve over time.

We can therefore see that algorithmic trading and trading done by the means of AI and ML are both widely researched topics. It is also to be noted that the results of these different research papers remain varied, and no general consensus amongst these mod-els truly exists. Overall, AI modmod-els can be seen as being better when it comes to the process of forecasting as opposed to plain static algorithms, as they have an inherent ability to adapt which is especially important due to the nature of the market that they operate in. It is also noted that the usage of multiple AI models is the preferred approach as opposed to a single forecast and lastly and most importantly it can be seen that hu-mans sometimes oppose to these systems, even if this does not make sense in terms of their overall performance.